page-cache: use the PAGEMAP_SCAN ioctl when it is available

Signed-off-by: Andrei Vagin <[email protected]>

criu/include/mem.h

@@ -7,6 +7,7 @@
 #include "pid.h"
 #include "proc_parse.h"
 #include "inventory.pb-c.h"
+#include "pagemap-cache.h"
 
 struct parasite_ctl;
 struct vm_area_list;
@@ -47,5 +48,6 @@ int open_vmas(struct pstree_item *t);
 int prepare_vmas(struct pstree_item *t, struct task_restore_args *ta);
 int unmap_guard_pages(struct pstree_item *t);
 int prepare_mappings(struct pstree_item *t);
-bool should_dump_page(VmaEntry *vmae, u64 pme);
+
+u64 should_dump_page(pmc_t *pmc, VmaEntry *vmae, u64 vaddr, u64 *ppme);
 #endif /* __CR_MEM_H__ */

criu/include/pagemap-cache.h

@@ -1,10 +1,12 @@
 #ifndef __CR_PAGEMAP_H__
 #define __CR_PAGEMAP_H__
 
+#include <stdbool.h>
 #include <sys/types.h>
 #include "int.h"
 
 #include "common/list.h"
+#include "pagemap_scan.h"
 
 struct vma_area;
 
@@ -15,9 +17,15 @@ typedef struct {
 	unsigned long start;		  /* start of area */
 	unsigned long end;		  /* end of area */
 	const struct list_head *vma_head; /* list head of VMAs we're serving */
+	int fd;				  /* file to read PMs from */
+
 	u64 *map;			  /* local buffer */
 	size_t map_len;			  /* length of a buffer */
-	int fd;				  /* file to read PMs from */
+
+	struct page_region *regs;	  /* buffer for the PAGEMAP_SCAN ioctl */
+	size_t regs_len;		  /* lenght of regs */
+	size_t regs_max_len;		  /* maximum length of regs */
+	size_t regs_idx;		  /* current index in the regs array */
 } pmc_t;
 
 #define PMC_INIT \
@@ -26,7 +34,7 @@ typedef struct {
 	}
 
 extern int pmc_init(pmc_t *pmc, pid_t pid, const struct list_head *vma_head, size_t size);
-extern u64 *pmc_get_map(pmc_t *pmc, const struct vma_area *vma);
+extern int pmc_get_map(pmc_t *pmc, const struct vma_area *vma, bool soft_dirty);
 extern void pmc_fini(pmc_t *pmc);
 
 #endif /* __CR_PAGEMAP_H__ */

criu/include/shmem.h

@@ -4,13 +4,14 @@
 #include "int.h"
 #include "common/lock.h"
 #include "images/vma.pb-c.h"
+#include "pagemap-cache.h"
 
 struct vma_area;
 
 extern int collect_shmem(int pid, struct vma_area *vma);
 extern int collect_sysv_shmem(unsigned long shmid, unsigned long size);
 extern int cr_dump_shmem(void);
-extern int add_shmem_area(pid_t pid, VmaEntry *vma, u64 *map);
+extern int add_shmem_area(pid_t pid, VmaEntry *vma, pmc_t *pmc);
 extern int fixup_sysv_shmems(void);
 extern int dump_one_memfd_shmem(int fd, unsigned long shmid, unsigned long size);
 extern int dump_one_sysv_shmem(void *addr, unsigned long size, unsigned long shmid);

criu/mem.c

@@ -99,38 +99,56 @@ static inline bool __page_in_parent(bool dirty)
 	return opts.track_mem && opts.img_parent && !dirty;
 }
 
-bool should_dump_page(VmaEntry *vmae, u64 pme)
+static bool should_dump_entire_vma(VmaEntry *vmae)
 {
 	/*
 	 * vDSO area must be always dumped because on restore
 	 * we might need to generate a proxy.
 	 */
-	if (vma_entry_is(vmae, VMA_AREA_VDSO))
+	if (vma_entry_is(vmae, VMA_AREA_VDSO)) {
 		return true;
-	/*
-	 * In turn VVAR area is special and referenced from
-	 * vDSO area by IP addressing (at least on x86) thus
-	 * never ever dump its content but always use one provided
-	 * by the kernel on restore, ie runtime VVAR area must
-	 * be remapped into proper place..
-	 */
-	if (vma_entry_is(vmae, VMA_AREA_VVAR))
-		return false;
-
-	/*
-	 * Optimisation for private mapping pages, that haven't
-	 * yet being COW-ed
-	 */
-	if (vma_entry_is(vmae, VMA_FILE_PRIVATE) && (pme & PME_FILE))
-		return false;
+	}
 	if (vma_entry_is(vmae, VMA_AREA_AIORING))
 		return true;
-	if ((pme & (PME_PRESENT | PME_SWAP)) && !__page_is_zero(pme))
-		return true;
 
 	return false;
 }
 
+/*
+ * should_dump_page returns vaddr if an addressed page has to be dumped.
+ * Otherwise, it returns an address that has to be inspected next.
+ */
+u64 should_dump_page(pmc_t *pmc, VmaEntry *vmae, u64 vaddr, u64 *ppme)
+{
+	if (pmc->regs) {
+		while (1) {
+			if (pmc->regs_idx == pmc->regs_len)
+				return vmae->end;
+			if (vaddr < pmc->regs[pmc->regs_idx].end)
+				break;
+			pmc->regs_idx++;
+		}
+		if (vaddr < pmc->regs[pmc->regs_idx].start)
+			return pmc->regs[pmc->regs_idx].start;
+		return vaddr;
+	} else {
+		u64 pme = pmc->map[PAGE_PFN(vaddr - pmc->start)];
+
+		*ppme = pme;
+
+		/*
+		 * Optimisation for private mapping pages, that haven't
+		 * yet being COW-ed
+		 */
+		if (vma_entry_is(vmae, VMA_FILE_PRIVATE) && (pme & PME_FILE))
+			return vaddr + PAGE_SIZE;
+		if ((pme & (PME_PRESENT | PME_SWAP)) && !__page_is_zero(pme))
+			return vaddr;
+
+		return vaddr + PAGE_SIZE;
+	}
+}
+
 bool page_is_zero(u64 pme)
 {
 	return __page_is_zero(pme);
@@ -164,25 +182,29 @@ static bool is_stack(struct pstree_item *item, unsigned long vaddr)
  * the memory contents is present in the parent image set.
  */
 
-static int generate_iovs(struct pstree_item *item, struct vma_area *vma, struct page_pipe *pp, u64 *map, u64 *off,
+static int generate_iovs(struct pstree_item *item, struct vma_area *vma, struct page_pipe *pp, pmc_t *pmc, u64 *pvaddr,
 			 bool has_parent)
 {
-	u64 *at = &map[PAGE_PFN(*off)];
-	unsigned long pfn, nr_to_scan;
+	unsigned long nr_scanned;
 	unsigned long pages[3] = {};
+	unsigned long vaddr;
+	bool dump_all_pages;
 	int ret = 0;
 
-	nr_to_scan = (vma_area_len(vma) - *off) / PAGE_SIZE;
-
-	for (pfn = 0; pfn < nr_to_scan; pfn++) {
-		unsigned long vaddr;
+	dump_all_pages = should_dump_entire_vma(vma->e);
+	
+	nr_scanned = 0;
+	for (vaddr = *pvaddr; vaddr < vma->e->end; vaddr += PAGE_SIZE, nr_scanned++) {
 		unsigned int ppb_flags = 0;
+		u64 pme = 0, next;
 		int st;
 
-		if (!should_dump_page(vma->e, at[pfn]))
+		/* If dump_all_pages is true, should_dump_page is called to get pme. */
+		next = should_dump_page(pmc, vma->e, vaddr, &pme);
+		if (!dump_all_pages && next != vaddr) {
+			vaddr = next - PAGE_SIZE;
 			continue;
-
-		vaddr = vma->e->start + *off + pfn * PAGE_SIZE;
+		}
 
 		if (vma_entry_can_be_lazy(vma->e) && !is_stack(item, vaddr))
 			ppb_flags |= PPB_LAZY;
@@ -194,7 +216,7 @@ static int generate_iovs(struct pstree_item *item, struct vma_area *vma, struct
 		 * page. The latter would be checked in page-xfer.
 		 */
 
-		if (has_parent && page_in_parent(at[pfn] & PME_SOFT_DIRTY)) {
+		if (has_parent && page_in_parent(pme & PME_SOFT_DIRTY)) {
 			ret = page_pipe_add_hole(pp, vaddr, PP_HOLE_PARENT);
 			st = 0;
 		} else {
@@ -214,9 +236,9 @@ static int generate_iovs(struct pstree_item *item, struct vma_area *vma, struct
 		pages[st]++;
 	}
 
-	*off += pfn * PAGE_SIZE;
 
-	cnt_add(CNT_PAGES_SCANNED, nr_to_scan);
+	*pvaddr = vaddr;
+	cnt_add(CNT_PAGES_SCANNED, nr_scanned);
 	cnt_add(CNT_PAGES_SKIPPED_PARENT, pages[0]);
 	cnt_add(CNT_PAGES_LAZY, pages[1]);
 	cnt_add(CNT_PAGES_WRITTEN, pages[2]);
@@ -356,12 +378,21 @@ static int generate_vma_iovs(struct pstree_item *item, struct vma_area *vma, str
 			     struct page_xfer *xfer, struct parasite_dump_pages_args *args, struct parasite_ctl *ctl,
 			     pmc_t *pmc, bool has_parent, bool pre_dump, int parent_predump_mode)
 {
-	u64 off = 0;
-	u64 *map;
+	bool soft_dirty = has_parent;
+	u64 vaddr;
 	int ret;
 
 	if (!vma_area_is_private(vma, kdat.task_size) && !vma_area_is(vma, VMA_ANON_SHARED))
 		return 0;
+	/*
+	 * In turn VVAR area is special and referenced from
+	 * vDSO area by IP addressing (at least on x86) thus
+	 * never ever dump its content but always use one provided
+	 * by the kernel on restore, ie runtime VVAR area must
+	 * be remapped into proper place..
+	 */
+	if (vma_entry_is(vma->e, VMA_AREA_VVAR))
+		return 0;
 
 	/*
 	 * To facilitate any combination of pre-dump modes to run after
@@ -390,7 +421,6 @@ static int generate_vma_iovs(struct pstree_item *item, struct vma_area *vma, str
 	 * pre-dump cycle has skipped processing & pagemap generation for
 	 * non-PROT_READ regions. So SPLICE mode throws error of missing
 	 * pagemap entry for encountered non-PROT_READ mapping.
-	 *
 	 * To resolve this, the pre-dump-mode is stored in current pre-dump's
 	 * inventoy file. This pre-dump mode is read back from this file
 	 * (present in parent pre-dump dir) as parent-pre-dump-mode during
@@ -421,15 +451,14 @@ static int generate_vma_iovs(struct pstree_item *item, struct vma_area *vma, str
 		has_parent = false;
 	}
 
-	map = pmc_get_map(pmc, vma);
-	if (!map)
+	if (pmc_get_map(pmc, vma, soft_dirty))
 		return -1;
 
 	if (vma_area_is(vma, VMA_ANON_SHARED))
-		return add_shmem_area(item->pid->real, vma->e, map);
-
+		return add_shmem_area(item->pid->real, vma->e, pmc);
+	vaddr = vma->e->start;
 again:
-	ret = generate_iovs(item, vma, pp, map, &off, has_parent);
+	ret = generate_iovs(item, vma, pp, pmc, &vaddr, has_parent);
 	if (ret == -EAGAIN) {
 		BUG_ON(!(pp->flags & PP_CHUNK_MODE));
 

criu/pagemap-cache.c

@@ -1,5 +1,6 @@
 #include <unistd.h>
 #include <fcntl.h>
+#include <sys/ioctl.h>
 
 #include "page.h"
 #include "pagemap-cache.h"
@@ -50,10 +51,19 @@ int pmc_init(pmc_t *pmc, pid_t pid, const struct list_head *vma_head, size_t siz
 	pmc->pid = pid;
 	pmc->map_len = PAGEMAP_LEN(map_size);
 	pmc->vma_head = vma_head;
+	pmc->regs_max_len = (map_size / PAGE_SIZE + 1) / 2;
+	pmc->regs_len = 0;
+	pmc->regs_idx = 0;
 
-	pmc->map = xmalloc(pmc->map_len);
-	if (!pmc->map)
-		goto err;
+	if (kdat.has_pagemap_scan) {
+		pmc->regs = xmalloc(pmc->regs_max_len * sizeof(struct page_region));
+		if (!pmc->regs)
+			goto err;
+	} else {
+		pmc->map = xmalloc(pmc->map_len);
+		if (!pmc->map)
+			goto err;
+	}
 
 	if (pagemap_cache_disabled)
 		pr_warn_once("The pagemap cache is disabled\n");
@@ -87,12 +97,7 @@ int pmc_init(pmc_t *pmc, pid_t pid, const struct list_head *vma_head, size_t siz
 	return -1;
 }
 
-static inline u64 *__pmc_get_map(pmc_t *pmc, unsigned long addr)
-{
-	return &pmc->map[PAGE_PFN(addr - pmc->start)];
-}
-
-static int pmc_fill_cache(pmc_t *pmc, const struct vma_area *vma)
+static int pmc_fill_cache(pmc_t *pmc, const struct vma_area *vma, bool soft_dirty)
 {
 	unsigned long low = vma->e->start & PMC_MASK;
 	unsigned long high = low + PMC_SIZE;
@@ -153,35 +158,60 @@ static int pmc_fill_cache(pmc_t *pmc, const struct vma_area *vma)
 	BUG_ON(pmc->map_len < size_map);
 	BUG_ON(pmc->fd < 0);
 
-	if (pread(pmc->fd, pmc->map, size_map, PAGEMAP_PFN_OFF(pmc->start)) != size_map) {
-		pmc_zap(pmc);
-		pr_perror("Can't read %d's pagemap file", pmc->pid);
-		return -1;
+	if (kdat.has_pagemap_scan && !soft_dirty) {
+		struct pm_scan_arg args = {
+			.size = sizeof(struct pm_scan_arg),
+			.flags = 0,
+			.start = pmc->start,
+			.end = pmc->end,
+			.vec = (long)pmc->regs,
+			.vec_len = pmc->regs_max_len,
+			.max_pages = 0,
+			.category_inverted = PAGE_IS_PFNZERO | PAGE_IS_FILE,
+			.category_mask = PAGE_IS_PFNZERO | PAGE_IS_FILE,
+			.category_anyof_mask = PAGE_IS_PRESENT | PAGE_IS_SWAPPED,
+			.return_mask = PAGE_IS_PRESENT | PAGE_IS_SWAPPED,
+		};
+		long ret;
+
+		ret = ioctl(pmc->fd, PAGEMAP_SCAN, &args);
+		if (ret == -1) {
+			pr_perror("PAGEMAP_SCAN");
+			pmc_zap(pmc);
+			return -1;
+		}
+		pmc->regs_len = ret;
+		pmc->regs_idx = 0;
+	} else {
+		if (pread(pmc->fd, pmc->map, size_map, PAGEMAP_PFN_OFF(pmc->start)) != size_map) {
+			pmc_zap(pmc);
+			pr_perror("Can't read %d's pagemap file", pmc->pid);
+			return -1;
+		}
 	}
 
 	return 0;
 }
 
-u64 *pmc_get_map(pmc_t *pmc, const struct vma_area *vma)
+int pmc_get_map(pmc_t *pmc, const struct vma_area *vma, bool soft_dirty)
 {
 	/* Hit */
 	if (likely(pmc->start <= vma->e->start && pmc->end >= vma->e->end))
-		return __pmc_get_map(pmc, vma->e->start);
+		return 0;
 
 	/* Miss, refill the cache */
-	if (pmc_fill_cache(pmc, vma)) {
+	if (pmc_fill_cache(pmc, vma, soft_dirty)) {
 		pr_err("Failed to fill cache for %d (%lx-%lx)\n", pmc->pid, (long)vma->e->start, (long)vma->e->end);
-		return NULL;
+		return -1;
 	}
-
-	/* Hit for sure */
-	return __pmc_get_map(pmc, vma->e->start);
+	return 0;
 }
 
 void pmc_fini(pmc_t *pmc)
 {
 	close_safe(&pmc->fd);
 	xfree(pmc->map);
+	xfree(pmc->regs);
 	pmc_reset(pmc);
 }